Customer churn features among the most challenging issues that adversely impact the growth strategy and revenue of a company. Thus, retaining existing customers remains a top priority for companies to sustain their operations and ensure business expansion. Artificial intelligence (AI) tools, like machine learning (ML) and deep learning (DL), have been utilized to successfully identify customer churn. Several studies have presented AI-based models for customer churn prediction. This article deliberates on the increasing importance of AI in customer churn prediction and some recent developments in this field.
Image credit: MUNGKHOOD STUDIO/Shutterstock
Why Customer Churn Prediction is Important?
Striking the perfect balance between customer retention and acquisition has become the crucial battleground in the age of intense competition. Thus, companies/businesses must implement appropriate mechanisms and systems for customer churn prediction and analysis as retaining existing customers is always an inexpensive approach.
Customer churn prediction involves predicting the possibility of a customer leaving a service or product of a company and moving over to its competitor in a highly competitive market. The purpose of churn prediction is to predict such customer behavior before the churn occurs and to implement actions to prevent the churn. This approach assists in establishing customer retention strategies, which primarily involves tracking loyal customers and their behavior.
Dissatisfaction with the existing consumer support and service system is one of the major reasons behind customer churn. The customer churn dataset is utilized to assess the marketing tendency of customers from large databases. Churn rate tracks lost customers within a particular time frame, shedding light on their reasons for leaving. Customer churn models can detect early churn signals and predict customers at risk of leaving the product/service.
Growing Role of AI
Companies are increasingly focusing on customer retention, which has driven the application and development of smarter ML models for prediction and analysis. In an ever-evolving market, service, product quality, and market fluctuations can result in customer loss. ML models can assist companies in identifying potential lost customers to reduce revenue losses significantly.
A variety of AI/ML algorithms/models have been explored in research for customer churn prediction. These include artificial neural networks (ANN), decision trees (DT) learning, regression analysis, logistic regression (LR), support vector machines (SVM), naïve Bayes (NB), linear discriminant analysis (LDA), and sequential pattern mining and market basket analysis.
For instance, tree-based algorithms, including XGBoost, GBM tree algorithm, random forest (RF), and DT, were applied for the customer churn prediction, and their performance was compared in a research. XGBoost outperformed all algorithms based on AUC accuracy in the comparative analysis. Feature selection, powered by optimization algorithms, can unlock even higher prediction accuracy.
Another study compared LR, NB, DT, and SVM to find the best tool for predicting customer churn. Additionally, the impact of boosting algorithms on classification accuracy was also observed. The results showed that SVM-POLY using AdaBoost had superior performance compared to other models. Feature selection strategies like uni-variate selection can lead to even higher classification accuracy.
Three ML approaches, including Bayesian networks, SVM, and neural networks, were adopted for customer churn prediction in another study. Additionally, the principal component analysis (PCA) reduced data dimensions in the feature selection process.
A churn model was constructed using different ML algorithms to assist telecom operators in predicting customers who are expected to churn. A comparative analysis determined the best model among the evaluated ML algorithms. Results displayed that RF combined with SMOTE-edited nearest neighbor (ENN) outperforms the other algorithms based on the F1-score, with the method achieving the output with 95% accuracy.
In another study, researchers presented a customer churn prediction methodology consisting of six phases for the telecom industry. Data preprocessing and feature analysis were performed in the first phase and second phase. Feature selection was performed using the gravitational search algorithm in the third phase.
Subsequently, the data was divided into the test set and the train set in a 20:80 ratio. The most common predictive models, like DT, RF, SVM, NB, and LR, were applied on the train set in the prediction process. Boosting and ensemble techniques were also employed to determine the impact on model accuracy.
Additionally, the k-fold cross-validation was applied to the train set to tune hyperparameters and avoid model overfitting. Eventually, the results obtained on the test set were evaluated using the AUC curve and confusion matrix. XGboost and Adaboost classifiers displayed the best accuracy of 80.8% and 81.71%, respectively. Moreover, both Adaboost and XGBoost classifiers achieved the highest AUC score of 84%.
DL-based Approaches
A study published in the Journal of Modelling in Management presented a model to assist e-businesses in predicting the churned users using ML. The objective of the distributed churn prediction model was to monitor customer behavior and make decisions accordingly.
The proposed model was based on the two-dimensional (2D) convolutional neural network (CNN), a DL technique. It possessed a layered architecture comprising two different phases, including the data load and preprocessing and the 2D CNN layer.
Additionally, the Apache Spark parallel and distributed framework was also employed for data processing in a parallel environment. Training data was obtained from Kaggle using Telco Customer Churn. Results validated the model's accuracy as the model attained a 0.963 accuracy score out of one. The validation and training loss was 0.004, which is extremely low.
Moreover, the confusion matric results demonstrated that the true-positive values and true-negative values were 95% and 94%, respectively. The false-negative and the false-positive values were 5% and 6%, respectively, indicating the model's effectiveness.
A study published in the journal Information Sciences Letters implemented the Deep-BP-ANN model to predict customer churn in the telecommunications industry. The model was implemented using two feature selection methods, including lasso regression and variance thresholding, and strengthened using the early stopping technique to prevent overfitting and stop training at the right time.
The efficiency of minimizing overfitting was compared between dropout and activity regularization strategies for two real-world datasets, including Cell2cell and IBM Telco. The random oversampling technique was utilized to balance both datasets. Ten-fold cross-validation and Holdout were utilized to determine the model efficiency. The model performance was compared with four ML techniques, including k-nearest neighbors (KNN), NB, LR, and XGBoost.
The proposed model displayed good performance with lasso regression for feature selection, activity regularization to minimize overfitting for both datasets, early stopping technique to select the epochs, and large numbers of neurons into the hidden and input layers.
Moreover, the Deep-BP-ANN model outperformed KNN, NB, LR, and XGBoost in predicting customer churn. The model’s accuracy also outperformed the existing DL techniques that employ holdout or 10-fold cross-validation for the same datasets.
Recent Developments
A paper published in Computational Intelligence and Neuroscience proposed a unique AI-based customer churn prediction model for telecommunication business markets (AICCP-TBM). The objective of the model is to regulate the existence of non-churners and churners in the telecom sector. AICCP-TBM employed a chaotic salp swarm optimization-based feature selection (CSSO-FS) method for the most suitable feature assortment, and a fuzzy rule-based classifier (FRC) was utilized to differentiate between client churners and non-churners.
Additionally, the quantum-behaved particle swarm optimization was used to select the FRC model’s membership functions to improve the classification performance. A benchmark customer churn prediction dataset was used to validate the performance of the AICCP-TBM model.
The AICCP-TBM model showed the best performance among the state-of-the-art models with the highest accuracy of 97.5% and the lowest accuracy of 94.33%. Moreover, the AICCP-TBM model also showed improved prediction performance in the simulation results.
In conclusion, AI/ML methods increase the effectiveness and accuracy of customer churn prediction substantially, allowing companies to take pre-emptive actions to minimize losses in a volatile market environment. However, the analytical predictions of common ML methods and business goals do not align effectively in several instances as these common models do not factor in the financial costs and benefits. Additionally, several factors that influence customer churn but are often not considered for customer churn prediction must be included to improve the prediction accuracy.
References and Further Reading
Lalwani, P., Mishra, M. K., Chadha, J. S., Sethi, P. (2022). Customer churn prediction system: a machine learning approach. Computing, 1-24. https://doi.org/10.1007/s00607-021-00908-y
Srinivasan, R., Rajeswari, D., Elangovan, G. (2023). Customer Churn Prediction Using Machine Learning Approaches. 2023 International Conference on Artificial Intelligence and Knowledge Discovery in Concurrent Engineering (ICECONF), 1-6. https://doi.org/10.1109/ICECONF57129.2023.10083813
Liang, Z. (2023). Predict Customer Churn based on Machine Learning Algorithms. Highlights in Business, Economics and Management, 10, 270-275. https://doi.org/10.54097/hbem.v10i.8051
Tariq, M. U., Babar, M., Poulin, M., Khattak, A. S. (2022). Distributed model for customer churn prediction using convolutional neural network. Journal of Modelling in Management, 17(3), 853-863. https://doi.org/10.1108/JM2-01-2021-0032
Faritha Banu, J., Neelakandan, S., Geetha, B. T., Selvalakshmi, V., Umadevi, A., Martinson, E. O. (2022). Artificial intelligence based customer churn prediction model for business markets. Computational Intelligence and Neuroscience. https://doi.org/10.1155/2022/1703696
Vafeiadis, T., Diamantaras, K., Sarigiannidis, G., Chatzisavvas, K. (2015). A comparison of machine learning techniques for customer churn prediction. Simulation Modelling Practice and Theory, 55, 1-9. https://doi.org/10.1016/j.simpat.2015.03.003
Fujo, S. W., Subramanian, S., Khder, M. A. (2022). Customer churn prediction in telecommunication industry using deep learning. Information Sciences Letters, 11(1), 24.https://digitalcommons.aaru.edu.jo/isl/vol11/iss1/24
De, S., Prabu, P., Paulose, J. (2021). Effective ML techniques to predict customer churn. 2021 Third international conference on inventive research in computing applications (ICIRCA),895-902. https://doi.org/10.1109/ICIRCA51532.2021.9544785